测试结果出现了很大问题 about crnn.caffe HOT 9 OPEN

一开始未做任何改动，完全按照步骤生成数据，然后训练报错，维度不匹配，根据相关问题里的解答，将reshape里的24和timestep的24都改为32后（24应该是宽度为96时的设置，而当前的图片宽度是128），可以顺利train起来，并在训练集和验证集都达到99的正确率。但是测试训练集图片时的输出都完全不对。需要修改两个地方：1.默认的验证码数据的单个字符仅包含0-9再加上空白标记总共11个，所以blank_label: 10，alphabet_size: 11，fc 层num_output: 11。2.训练时图片数据是没有归一化到0-1的，而recognition.cpp中sample_resized.convertTo(sample_float, CV_32FC3)将图片像素值归一化到0-1，将其改为sample_resized.convertTo(sample_float, CV_32FC3, 1/255.0)。

from crnn.caffe.

@plastic0313 我也有这样的问题，请问您的解决了么

from crnn.caffe.

@plastic0313 @greatgeekgrace 因为的自己的分类个数是74，你要参考你自己的分类个数，更改generate_dataset.py num_output alphabet_size等参数。

from crnn.caffe.

@yalecyu 好的，非常感谢~~~目前预测的captcha图片（图片为0642）结果：
74 74 0 74 6 74 4 74 2 74 74 74 74 74 74 74 74 74 74 74 - - - -
看起来结果有24个字符（上面的结果一个一个数的），可是在crnn.prototxt和deploy.prototxt中num_output设置的是75，是怎么回事呢。而且还出现了-符号？

from crnn.caffe.

@xijunjun 对，主要注意的是，因为另一个OCR的项目，我更改了prototxt的配置，没有用生成数据集验证是否维度匹配。另一个需要注意的就是0-1和0-255，但是没有验证过，只是README里面给出提示。有时间补了这些坑。

from crnn.caffe.

@xijunjun 哪儿可以知道训练的时候是归一化的？convertTo(sample_float, CV_32FC3, 1/255.0)这个函数的作用不是归一化吗？

from crnn.caffe.

@dingtao1 我是看了下数据制作代码和数据输入层参数

from crnn.caffe.

Hi, sorry for posting on an old discussion, but i need some help or hints as I can't seem to get consistent and correct results for my own trained crnn model after following all the steps. I ported the Linux code and compiled this on Windows and Visual Studio 2017 compiler. I managed to compile the codes successfully after making some changes, but I supposed this shouldn't affect the results.

1. First I generated dataset using generate_captcha.py. Total image size is 50,000.

2. Then execute generate_dateset.py.
   IMAGE_WIDTH, IMAGE_HEIGHT = 128, 32.
   Training size = 40,000 and Test size = 10,000.

3. In my crnn.prototxt, I changed batch size to 50 to cater for my GPU which only have 2 MB memory. I changed the following as well:
   layer {
   name: "reshape"
   type: "Reshape"
   bottom: "conv6"
   top: "reshape"
   reshape_param {
   shape {
   #nc(w*h)
   dim: 50
   dim: 512
   dim: 32
   }
   }
   }
   layer {
   name: "indicator"
   type: "ContinuationIndicator"
   top: "indicator"
   continuation_indicator_param {
   time_step: 32
   batch_size: 50
   }
   }
   layer {
   name: "ctc_loss"
   type: "CtcLoss"
   bottom: "fc1"
   bottom: "label"
   top: "ctc_loss"
   loss_weight: 1.0
   ctc_loss_param {
   blank_label: 10
   alphabet_size: 11
   time_step: 32
   }
   }
   layer {
   name: "accuracy"
   type: "LabelsequenceAccuracy"
   bottom: "premuted_fc"
   bottom: "label"
   top: "accuracy"
   labelsequence_accuracy_param {
   blank_label: 10
   }
   }
   I managed to get over 0.95 accuracy for both test and train data. Loss seems to be on the low side as well (0.00x).

4. Next, I change the deploy.prototxt:
   name: "crnn"
   layer {
   name: "data"
   type: "Input"
   top: "data"
   input_param {shape:{dim:1 dim:3 dim:32 dim:128}}
   }
   layer {
   name: "reshape"
   type: "Reshape"
   bottom: "conv6"
   top: "reshape"
   reshape_param {
   shape {
   #nc(w*h)
   dim: 1
   dim: 512
   dim: 32
   }
   }
   }
   layer {
   name: "indicator"
   type: "ContinuationIndicator"
   top: "indicator"
   continuation_indicator_param {
   time_step: 32
   batch_size: 1
   }
   }
   layer {
   name: "fc1"
   type: "InnerProduct"
   bottom: "lstm2"
   top: "fc1"
   param {
   lr_mult: 1
   decay_mult: 1
   }
   param {
   lr_mult: 2
   decay_mult: 0
   }
   inner_product_param {
   num_output: 11
   axis: 2
   weight_filler {
   type: "xavier"
   }
   bias_filler {
   type: "constant"
   value: 0
   }
   }
   }

5. I also amend the recognition.cpp to include the normalization:
   if (num_channels_ == 3)sample_resized.convertTo(sample_float, CV_32FC3, 1.f/255);
   else sample_resized.convertTo(sample_float, CV_32FC1, 1.f/255);

for the output, when i run the recognition exe such as below:
recognition D:\ImageProc\ImgDataset\Data\Captcha\49998-7959.png D:\Lib\caffecrnn\examples\crnn\deploy.prototxt D:\Lib\caffecrnn\examples\crnn\model\crnn_captcha_iter_3600.caffemodel
i can't get consistent results from the model each time, and i can't get an accurate output as well:
Output that I get if i run it for three times:
8 9 8 8 8 8 8 8 8 8 8 8 8 8 8 4 4 4 4 4 1 1 1 - 1 1 1 0 0 - 1 2

6 6 - - 9 9 - - - - - - - - - - - - 6 6 6 6 6 6 6 6 6 6 6 6 3 3

1 1 2 1 1 8 8 8 8 1 - 1 1 1 1 1 1 1 1 7 7 7 7 7 7 7 7 7 7 7 6 1

Anyone have any hints or detected where I have a mistake? Anyone managed to get accurate output from the trained model?

Please help! Thank you.

from crnn.caffe.

（数字+英文字母）测试图从BGR转RGB可以解决训练过程中测试准确率很高，但是cpp_recognition输出结果不对的问题！！！
#～～～～～～～～～～～～～～～～～～原因如下～～～～～～～～～～～～～～～～～～～～#
我们做数据的时候：img = caffe.io.load_image(os.path.join(img_path, image))
caffe.io里面：img = skimage.img_as_float(skimage.io.imread(filename, as_grey=not color)).astype(np.float32)
问题所在：cv2的存储格式是BGR，而skimage的存储格式是RGB（recognition.cpp里面的读图是用opencv,使用cv::cvtColor(resizeimg, resizeimg, cv::COLOR_BGR2RGB);）

from crnn.caffe.